Online/Streaming Inference of New Nodes with DGL

Hello, I am trying to build an application based on dgl that would use GNNs for inference in an online/streaming manner, i.e., node classification on new nodes coming to the graph one by one

Since I need mutable graph, I am working with DGLHeterograph instead of any GraphBolt-based representation
Given this setup, how to best handle the inference for a set of given nodes? During each iteration, I want to compute embedding only for a given node, so I want something resembling dgl DataLoader with MFGs, but I dont think creating a new DataLoader instance with one index on every iteration would be the right way

Any suggestions or guidelines?